A neural-mechanistic hybrid approach improving the predictive power of genome-scale metabolic models